Use of Multiobjective Optimization for Data Clustering

Abstract

This last chapter deals with some multiobjective clustering techniques based on symmetry. Three different clustering techniques are discussed. These use the multiobjective simulated annealing technique, AMOSA, as the underlying optimization strategy. The first one (MOPS) partitions the data into a pre-specified number of clusters, while the second one (VAMOSA) can automatically determine the number of clusters. These are multiobjective extensions of GAPS and VGAPS respectively. Results show that both MOPS and VAMOSA outperform GAPS and VGAPS respectively for most of the data sets. Finally a generalized clustering technique, named GenClustMOO, is described in this chapter, which is well suited to detect the appropriate partitioning from data sets having either point-symmetric or well-separated clusters. Here, multiple seed points are used to encode a particular cluster. Three cluster validity indices, namely one based on Euclidean distance, another based on the point symmetry distance, and the third based on a new definition of connectivity between the points, are optimized simultaneously. Concepts of relative neighborhood graph are utilized to compute the connectivity index. Extensive experimental results illustrating the effectiveness of the multiobjective clustering techniques over the single objective approaches are also presented for several artificial and real-life data sets.